“One of the most consequential dimensions of AI that remains comparatively under-examined is its environmental footprint and the justice implications that follow.” This is how the latest report from the United Nations University Institute for Water, Environment and Health (UNU-INWEH), edited by Kaveh Madani, who was recently interviewed by our editorial staff, begins. The data emerging from the report, Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints, is striking. Of course, we knew that the sector had shown worrying hyper-growth, and few had taken an interest in the real impacts. But now the figures have been analysed and processed by one of the United Nations’ most influential working groups.
Starting with the economic figures: spending on infrastructure for training and computing will exceed $2.5 trillion in 2026. An outlay that mainly goes towards energy. If data centres, the physical backbone of AI, were a nation, their estimated consumption of 448 terawatt-hours (448 billion kWh) in 2025 would place them eleventh in the world, roughly on par with France. And by 2040, a further 40% increase is expected, accounting for 3% of total global electricity consumption. To put it in perspective, this is equivalent to the consumption of 1.3 billion people in sub-Saharan Africa. For five years. Shall we calculate this in terms of CO₂e emissions? AI would have as much impact on the climate as the emissions of the entire United Kingdom in 2025.
However, it is not just a question of greenhouse gas emissions and energy costs. The land-use impact of the necessary IT and energy infrastructure is projected to exceed 14,000 km² by the end of the decade – roughly the size of Northern Ireland. When it comes to water, meanwhile, the figures are even more staggering: an estimated 9.3 trillion litres of water consumed by data centres. A quantity sufficient to meet the drinking water needs of the Earth’s 8.1 billion inhabitants for around a year and a half. So much for turning off the tap when we brush our teeth. Sure, there are circular water management systems, but the report warns that even when some of the water extracted is returned, “large-scale withdrawals can put a strain on aquifers and river systems, especially in arid regions or where aquifers are already depleted”.
The problem, however, does not lie in the technology but in how we use it. While AI will be essential in managing water and electricity networks, driving innovation across numerous fields (from healthcare to clean tech) and improving the efficiency of industrial processes, the primary use made of these supercomputers today is decidedly trivial.
Estimates suggest that ChatGPT alone processes around 2.5 billion requests per day. Based on a modest estimate of 0.42 Wh per text request, this equates to approximately 383 GWh of electricity consumed annually. The corresponding annual water footprint would be equivalent to the minimum domestic water requirements of around 500,000 people in sub-Saharan Africa, while the land footprint exceeds 800 football pitches. But the problem is that the vast majority of people ask the AI to polish an email, suggest a recipe for dinner, tell them who Julius Caesar was, or advise on how to invest €500 in cryptocurrencies. Or, worse still, they use it to hear Grok utter a string of unspeakable obscenities or generate fake nude images of schoolmates.
All things we could easily do using any search engine (excluding the above-mentioned obscenities, of course). According to the report, a traditional search consumes around 0.3 Wh. A generative search powered by artificial intelligence consumes up to 3 Wh – ten times as much. While agentic AI can certainly improve work processes (even the editorial team at Renewable Matter uses AI for business management, reporting or analysing large data matrices), image and video generative AI is often used by trashy social media influencers, trolls and keyboard warriors, or despicable marketing hucksters, to produce what can bluntly be described as “slop”: dancing dogs, fake Star Wars remakes, AI-generated porn, images of President Trump healing the sick like Jesus Christ.
A single high-resolution AI-generated video can require over 415 Wh, resulting in higher energy consumption than the creation of hundreds of AI-generated images. When resolution and frame rate are taken into account, energy requirements increase squarely (doubling production quadruples energy consumption). And, with the proliferation of videos on the most popular platforms, this is rapidly becoming an infrastructure-level problem.
And, last but not least, let’s talk about materials: a huge amount is needed to build data centres, and just as much is generated as waste. “At the end of their life cycle, e-waste handled inappropriately can expose the most vulnerable communities to hazardous substances. By 2030, artificial intelligence infrastructure could generate up to 2.5 million tonnes of electronic waste per year, roughly equivalent to disposing of 250 Eiffel Towers every year,” explains the UNU report.
There is therefore an urgent need to reflect on what we should actually do with AI, not only in terms of its social and psychological impacts but also its energy and environmental impacts, which are just as pressing. “The future of artificial intelligence should not be assessed solely on the basis of what machines are capable of doing but also on humanity’s ability to employ those powers within the planet’s limits. Although often described as immaterial and virtual, the reality of AI is profoundly physical. Behind every command, image or video lies an ever-expanding infrastructure comprising energy systems, water extraction, land use, mining and electronic waste. This report is a call to highlight these hidden environmental costs before they become unmanageable,” says Kaveh Madani in the press release launching the report. An urgent appeal that the two countries currently home to 90% of AI infrastructure (the US and China) must take seriously. Even if it is clear that only one nation will do its homework properly. And it is not the country that speaks English in schools.
Cover: Envato Elements
“One of the most consequential dimensions of AI that remains comparatively under-examined is its environmental footprint and the justice implications that follow.” This is how the latest report from the United Nations University Institute for Water, Environment and Health (UNU-INWEH), edited by Kaveh Madani, who was recently interviewed by our editorial staff, begins. The data emerging from the report, Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints, is striking. Of course, we knew that the sector had shown worrying hyper-growth, and few had taken an interest in the real impacts. But now the figures have been analysed and processed by one of the United Nations’ most influential working groups.
Starting with the economic figures: spending on infrastructure for training and computing will exceed $2.5 trillion in 2026. An outlay that mainly goes towards energy. If data centres, the physical backbone of AI, were a nation, their estimated consumption of 448 terawatt-hours (448 billion kWh) in 2025 would place them eleventh in the world, roughly on par with France. And by 2040, a further 40% increase is expected, accounting for 3% of total global electricity consumption. To put it in perspective, this is equivalent to the consumption of 1.3 billion people in sub-Saharan Africa. For five years. Shall we calculate this in terms of CO₂e emissions? AI would have as much impact on the climate as the emissions of the entire United Kingdom in 2025.
However, it is not just a question of greenhouse gas emissions and energy costs. The land-use impact of the necessary IT and energy infrastructure is projected to exceed 14,000 km² by the end of the decade – roughly the size of Northern Ireland. When it comes to water, meanwhile, the figures are even more staggering: an estimated 9.3 trillion litres of water consumed by data centres. A quantity sufficient to meet the drinking water needs of the Earth’s 8.1 billion inhabitants for around a year and a half. So much for turning off the tap when we brush our teeth. Sure, there are circular water management systems, but the report warns that even when some of the water extracted is returned, “large-scale withdrawals can put a strain on aquifers and river systems, especially in arid regions or where aquifers are already depleted”.
The problem, however, does not lie in the technology but in how we use it. While AI will be essential in managing water and electricity networks, driving innovation across numerous fields (from healthcare to clean tech) and improving the efficiency of industrial processes, the primary use made of these supercomputers today is decidedly trivial.
Estimates suggest that ChatGPT alone processes around 2.5 billion requests per day. Based on a modest estimate of 0.42 Wh per text request, this equates to approximately 383 GWh of electricity consumed annually. The corresponding annual water footprint would be equivalent to the minimum domestic water requirements of around 500,000 people in sub-Saharan Africa, while the land footprint exceeds 800 football pitches. But the problem is that the vast majority of people ask the AI to polish an email, suggest a recipe for dinner, tell them who Julius Caesar was, or advise on how to invest €500 in cryptocurrencies. Or, worse still, they use it to hear Grok utter a string of unspeakable obscenities or generate fake nude images of schoolmates.
All things we could easily do using any search engine (excluding the above-mentioned obscenities, of course). According to the report, a traditional search consumes around 0.3 Wh. A generative search powered by artificial intelligence consumes up to 3 Wh – ten times as much. While agentic AI can certainly improve work processes (even the editorial team at Renewable Matter uses AI for business management, reporting or analysing large data matrices), image and video generative AI is often used by trashy social media influencers, trolls and keyboard warriors, or despicable marketing hucksters, to produce what can bluntly be described as “slop”: dancing dogs, fake Star Wars remakes, AI-generated porn, images of President Trump healing the sick like Jesus Christ.
A single high-resolution AI-generated video can require over 415 Wh, resulting in higher energy consumption than the creation of hundreds of AI-generated images. When resolution and frame rate are taken into account, energy requirements increase squarely (doubling production quadruples energy consumption). And, with the proliferation of videos on the most popular platforms, this is rapidly becoming an infrastructure-level problem.
And, last but not least, let’s talk about materials: a huge amount is needed to build data centres, and just as much is generated as waste. “At the end of their life cycle, e-waste handled inappropriately can expose the most vulnerable communities to hazardous substances. By 2030, artificial intelligence infrastructure could generate up to 2.5 million tonnes of electronic waste per year, roughly equivalent to disposing of 250 Eiffel Towers every year,” explains the UNU report.
There is therefore an urgent need to reflect on what we should actually do with AI, not only in terms of its social and psychological impacts but also its energy and environmental impacts, which are just as pressing. “The future of artificial intelligence should not be assessed solely on the basis of what machines are capable of doing but also on humanity’s ability to employ those powers within the planet’s limits. Although often described as immaterial and virtual, the reality of AI is profoundly physical. Behind every command, image or video lies an ever-expanding infrastructure comprising energy systems, water extraction, land use, mining and electronic waste. This report is a call to highlight these hidden environmental costs before they become unmanageable,” says Kaveh Madani in the press release launching the report. An urgent appeal that the two countries currently home to 90% of AI infrastructure (the US and China) must take seriously. Even if it is clear that only one nation will do its homework properly. And it is not the country that speaks English in schools.
Cover: Envato Elements
